A Tsinghua-affiliated AI company is pursuing a development path strikingly similar to that of leading global AI laboratories. On April 20, UBS analysts including Wei Xiong issued a nearly 40-page research report initiating coverage on Beijing-based AI firm KNOWLEDGE ATLAS with a "Buy" rating and a target price of HK$1,160.
The report's core conclusion is straightforward: KNOWLEDGE ATLAS's model development and commercialization strategy closely mirrors that of top global AI firm Anthropic—leading analysts to position it as "China's Anthropic."
Why the comparison to Anthropic? Anthropic is one of Silicon Valley's most prominent AI giants, focusing not on flashy video generation but intensely specializing in one area: enabling AI to write code. UBS asserts that KNOWLEDGE ATLAS aligns with Anthropic in both technical approach and revenue generation methods.
First, strategic focus is highly aligned. Both companies have chosen programming capabilities as their breakthrough point, reasoning that programming tasks yield verifiable results and quantifiable value, representing the shortest path for AI to evolve from "assistance" to "execution." Anthropic bets on coding capabilities at the model level, emphasizing long-horizon tasks—the ability of AI to complete complex engineering tasks lasting several hours without human intervention. KNOWLEDGE ATLAS has adopted an almost identical path, launching its AI programming assistant CodeGeeX as early as September 2022, making it one of China's first movers in AI programming. According to tests by METR, Anthropic's Claude Opus 4.6 can handle approximately 12 hours of human-equivalent task duration at a 50% success rate, while KNOWLEDGE ATLAS's GLM-5.1 achieves about 8 hours, ranking first among global open-source models.
Second, the performance gap between their models continues to narrow. Data from Artificial Analysis shows that as of April 17, 2026, KNOWLEDGE ATLAS's flagship model GLM-5.1 ranks seventh globally in the comprehensive intelligence index with a score of 51.4, just 2 points behind Anthropic's Claude Opus 4.6 at 53 points. In the SWE-bench Pro benchmark, which simulates real-world code engineering tasks, GLM-5.1 scores 58.4 points, ranking second globally behind Anthropic's Claude Mythos Preview and ahead of GPT-5.4.
Third, monetization trajectories are similar, but KNOWLEDGE ATLAS demonstrates faster Annual Recurring Revenue growth and better cost efficiency. In December 2025, KNOWLEDGE ATLAS's open platform ARR was $39 million. By March 2026, this figure reached $250 million—a 6.4-fold increase in just four months. By comparison, Anthropic took approximately nine months to achieve a similar growth multiple during its early stage. While Anthropic recently grew from $9 billion to $30 billion in just four months, this represents a 3.3-fold increase. Both companies monetize rapidly, but KNOWLEDGE ATLAS shows superior momentum. The report notes: "Compared to the global leader, KNOWLEDGE ATLAS exhibits a steeper ARR growth curve, despite starting from a smaller base."
Simultaneously, KNOWLEDGE ATLAS offers better value. GLM-5.1's comprehensive pricing is approximately $2 per million tokens, while the similarly capable Claude Opus 4.6 costs about $9—making KNOWLEDGE ATLAS's pricing roughly 22% of Anthropic's. This higher cost efficiency suggests room for future price increases as the model's capabilities advance.
In summary, both companies focus on programming rather than multimodality at the model level, develop proprietary programming agents at the engineering level, and employ capability-based pricing in commercialization, where higher capabilities command premium prices without dampening demand.
Programming represents the core battleground. Understanding KNOWLEDGE ATLAS requires recognizing why it treats programming capability as its primary strategic pivot. The report's logic is that programming ability constitutes the critical threshold for AI's entry into real enterprise scenarios. Code-writing capability意味着 AI can execute complex multi-step tasks and replace genuine engineering workflows, thereby significantly enhancing commercial value. KNOWLEDGE ATLAS has pursued this path for nearly four years, evolving from CodeGeeX 1 in 2022 to GLM-5.1 in April 2026—completing the transition from a "code completion tool" to a system capable of independently handling complex engineering tasks for 8-hour durations.
Specifically, in METR's long-horizon task testing, GLM-5.1 achieves approximately 8 hours of task completion span at 50% success rate, ranking first among global open-source models. In Code Arena's agent programming task rankings, GLM-5.1 places third globally.
External benchmarks and industry research indicate that GLM series models have become a "preferred choice" for enterprise clients in programming-related tasks, demonstrating strong user recognition.
Pricing power provides the most direct validation of commercialization capability. In February 2026, KNOWLEDGE ATLAS raised prices for its Coding Plan by 30% while tightening usage limits and model selection. Surprisingly, according to OpenRouter data, total token usage surged approximately threefold month-over-month in March 2026. More notably, GLM-5.0—then the highest-priced model in KNOWLEDGE ATLAS's portfolio—recorded the highest usage share among all models that month. The report interprets this as demonstrating "user willingness to pay for high-performance models and demand resilience, with price increases not suppressing usage."
KNOWLEDGE ATLAS's pricing evolution follows two parallel tracks: model upgrade-based increases—GLM-5.0's input prices rose 100% and output prices 125% compared to GLM-4.7—and direct adjustments like the 30% Coding Plan increase.
Compared to Anthropic, KNOWLEDGE ATLAS maintains significant cost advantage. GLM-5.1 delivers 97% of Claude Opus 4.6's comprehensive intelligence score at approximately 22% of the price, indicating substantial upward pricing potential as capabilities improve.
KNOWLEDGE ATLAS's research foundation stems from Tsinghua University. All five founders are Tsinghua alumni, with Tsinghua's asset management company holding a 3.53% stake. The company maintains deep collaboration with Tsinghua's Knowledge Engineering Research Group. The R&D team exceeds 800 people, with researchers comprising 74% of staff—both figures exceeding those of comparable emerging AI labs in China. The research team had published approximately 500 high-impact papers by June 30, 2025.
Technological innovations highlighted in the report include dynamic sparse attention mechanisms, the "Slime" asynchronous reinforcement learning framework, and native agent integration design.
KNOWLEDGE ATLAS employs a dual revenue model: private deployment and open platform. In 2025, private deployment accounted for 74% of total revenue, while the open platform API contributed 26%.
Private deployment serves as a stable foundation, primarily targeting government and large enterprises. Clients include nine of China's top ten internet companies and government-backed entities. In 2025, the company won 57 major model projects in public service with total contract value of ¥254 million, ranking fifth among domestic large model providers. Gross margins for private deployment maintain a healthy 50%-70% range, though days sales outstanding increased from 107 to 153 days in 2025, with accounts receivable growing from ¥100 million to ¥339 million.
The open platform represents the future growth engine. The programming package launched in September 2025, followed by the "Claw package" for agent framework needs in March 2026. The Claw package attracted 100,000 new users within two days and reached 400,000 within 20 days. Total platform users hit 4 million by March 2026. Paid token consumption surged 15-fold over six months. UBS projects open platform revenue will grow at a 470% CAGR from 2025's ¥190 million to ¥6.188 billion in 2027, with total revenue growing at a 231% CAGR from ¥724 million to ¥7.941 billion.
However, KNOWLEDGE ATLAS remains loss-making, with projected net losses of ¥5.157 billion in 2026 and ¥4.747 billion in 2027, potentially achieving profitability by 2029. Major downside risks include intensified industry competition, potential client loss to self-developed models by major internet platforms, computing power constraints, and geopolitical factors.
Comments